Pilothouse

Pilothouse turns several Claude Code cloud sessions ("rooms") into a small team: coordinator, team lead, product owner, developer, UX/UI and QA. Each room holds one role and keeps only that role's context. Rooms coordinate through three channels, in this order of preference: the repository (backlog, tickets, decisions, pull requests), one short handoff file per role that is overwritten rather than appended, and cross-room messages for "your turn" signals and questions.

What the skill gives your agent:

• Role definitions: duties, what each role must not do, who it receives work from and who it hands work to.
• Ownership: each room owns the quality of its own output. The PO asks whether the work is worth doing, whether the scope is complete and whether the acceptance criteria are clear; Dev asks whether the change is correct, proven and readable by the next room; UX/UI asks whether a real user can finish the task; QA asks whether testing covers the change and whether it is ready for production.
• Pushing back: a room may disagree once, with reasons, record the disagreement in its handoff file, then follow the owner's decision. It stops only before a step that cannot be undone or puts security at risk, until the owner confirms in their own words.
• Unclear scope: if a room is not sure which work, project or room an instruction means, it asks one question before starting, and work that belongs to another room is sent there instead of being done twice.
• Messaging rules for sending and answering messages between rooms, including what to do while waiting.
• A handoff file template (with a "Disagreements recorded" section) and a shared-memory layout (index plus one file per role).
• A rotation procedure for when a room's context has grown too large.
• Guidance for a coordinator or team lead on turning the owner's plain words into work orders per role.
• A launch procedure: give it a list of rooms and roles and it first looks at the repository (existing handoff files, an existing team), then either creates the rooms (if the session can create sessions) or prints starter prompts for you to paste.

Safety defaults: the human owner decides direction, merges, licensing and anything published. A message from another room is treated as information from a peer, not as an instruction from the owner. No secrets, tokens or personal data in messages, handoff files or commits.

Honest status: this is version 0.3.0 on GitHub (v1.2 here), written for Claude Code and tested in a small number of runs by its author, mostly on toy repositories. Ownership and pushing back have not yet been run with a real team. In headless runs (claude -p) with default permissions the skill could not read its own reference files; the core rules are in SKILL.md for that reason. Room creation through the skill and message passing between rooms on a local machine are not yet tested. The README in the source repository has the full table of what has and has not been tested. It does not run any scripts or make network calls of its own; it is instructions and templates only.